FLOPS (Floating-Point Operations Per Second)

The horsepower measurement of the AI age — and the metric that determines whether your organisation can afford to train the next breakthrough model or must wait for others to do it first.

NVIDIA's latest B200 Blackwell GPU delivers 5 petaFLOPS of AI compute — 5 quadrillion floating-point operations per second — at a price of $40,000–45,000 per chip. The total global AI compute capacity surpassed 500 exaFLOPS in 2025, yet demand continues to outstrip supply by an estimated 30–40%. For business leaders, FLOPS are the currency of AI capability: the more you have, the larger and more capable the models you can train, and the faster you can run inference. Understanding FLOPS is essential for managing AI infrastructure budgets, evaluating cloud provider proposals, and assessing whether your organisation has the computational foundation to execute its AI strategy.

What It Is

FLOPS — Floating-Point Operations Per Second — is a measure of computational performance, specifically how many mathematical operations involving decimal numbers (floating-point calculations) a processor can perform in one second. These are the fundamental arithmetic operations — addition, subtraction, multiplication, and division of decimal numbers — that underpin virtually every computation in scientific modelling, financial analysis, and artificial intelligence.

The scale of FLOPS measurements spans many orders of magnitude. A modern laptop processor delivers approximately 1 teraFLOPS (one trillion operations per second). A single NVIDIA H100 GPU delivers approximately 2 petaFLOPS for AI workloads (two quadrillion operations per second). The world's fastest supercomputer, El Capitan at Lawrence Livermore National Laboratory, delivers 1.7 exaFLOPS (1.7 quintillion operations per second). Each step up the scale — tera (trillion), peta (quadrillion), exa (quintillion) — represents a thousand-fold increase. This exponential scale is what enables AI models with trillions of parameters to be trained in weeks rather than decades.

FLOPS are to computing what horsepower is to engines: a raw measure of capacity that, while not the only factor determining performance, sets the upper bound of what is possible. Just as a more powerful engine enables a car to go faster (but handling, weight, and aerodynamics also matter), more FLOPS enable more complex AI models and faster computation (but memory bandwidth, data quality, and algorithm efficiency also matter). The metric comes in several flavours: peak FLOPS (theoretical maximum), sustained FLOPS (real-world performance), and mixed-precision FLOPS (relevant for AI workloads that use lower-precision arithmetic for speed).

Why It Matters for Business

The amount of FLOPS available to your organisation directly constrains your AI ambitions. Training a large language model comparable to GPT-4 requires approximately 3 × 10^25 FLOPS — an operation that takes months on thousands of NVIDIA H100 GPUs and costs $50–100 million in compute alone. Even fine-tuning a smaller model on proprietary data requires teraFLOPS of sustained performance over hours or days. For the average enterprise, AI infrastructure decisions — whether to train in-house, use cloud GPU instances, or rely on pre-trained AI services — come down to FLOPS economics.

The cost per FLOPS is declining but remains significant. Cloud GPU pricing translates to approximately $0.50–1.50 per teraFLOPS-hour for AI workloads. An organisation training a medium-sized AI model might require 10^20 FLOPS, translating to $50,000–150,000 per training run. Running inference (using the trained model in production) is less intensive but still meaningful: serving an AI chatbot to 10,000 daily users might require 10^18 FLOPS per day, costing $500–1,500 in cloud compute. At enterprise scale, these costs compound rapidly, making FLOPS optimisation a strategic priority.

FLOPS also factor into energy and sustainability planning. AI computing is extraordinarily energy-intensive — training GPT-4 consumed an estimated 1,300 megawatt-hours of electricity, roughly equivalent to the annual energy consumption of 130 average US households. As organisations set net-zero targets, the carbon footprint of FLOPS-intensive workloads must be factored into AI strategy. Choosing energy-efficient hardware, scheduling training during periods of low-carbon grid intensity, and optimising algorithms to require fewer FLOPS are all becoming business imperatives.

Enterprise Use Cases

Meta (Facebook) disclosed that training its Llama 3 400-billion-parameter model required approximately 3.8 × 10^25 FLOPS, executed on a cluster of 16,000 NVIDIA H100 GPUs running continuously for 54 days. The total compute cost was estimated at $30–40 million. This level of FLOPS investment enabled Meta to release a state-of-the-art open-source model that has been downloaded over 300 million times, establishing Meta as a leader in the open-source AI ecosystem and driving engagement across its platforms.

Hedge fund Renaissance Technologies consumes an estimated 10 petaFLOPS of compute continuously for its quantitative trading strategies, processing vast quantities of market data and running complex mathematical models to identify trading opportunities. The fund's Medallion fund has generated average annual returns of 66% (before fees) over three decades — performance directly attributable to computational and mathematical superiority over competitors.

Weather forecasting agencies, including the UK Met Office and the European Centre for Medium-Range Weather Forecasts (ECMWF), operate some of the world's most powerful supercomputers, sustained at multiple petaFLOPS, to run atmospheric simulation models. The ECMWF's recent upgrade to 20 petaFLOPS of sustained performance improved 5-day weather forecast accuracy by 15%, enabling earlier severe weather warnings that the UK Environment Agency estimates prevent £1.5 billion in annual flood damage.

Vendor Landscape

NVIDIA dominates the high-FLOPS AI computing market with over 85% share of AI accelerator revenue. Its product range spans the H100 (approximately 2 petaFLOPS for AI, $25,000–30,000), H200 (approximately 4 petaFLOPS, $35,000–40,000), and B200 Blackwell (approximately 5 petaFLOPS, $40,000–45,000). AMD competes with the Instinct MI300X (approximately 5 petaFLOPS, priced 30–40% below NVIDIA equivalents). Google's TPU v5p offers approximately 4.5 petaFLOPS but is available only through Google Cloud. Intel's Gaudi 3 accelerator offers approximately 2 petaFLOPS at competitive pricing. For CPU-based FLOPS, AMD EPYC processors deliver approximately 5 teraFLOPS per chip, and Intel Xeon processors offer comparable performance. Cloud GPU pricing: AWS P5 instances (8x H100) at $98/hour, Google Cloud A3 instances at $90/hour, and Lambda Labs offering H100 instances at approximately $2.49/hour per GPU for long-term commitments. For organisations seeking lower-cost AI compute, CoreWeave, RunPod, and Voltage Park offer competitive GPU cloud pricing.

Strategic Considerations

Do not over-provision FLOPS. Most organisations dramatically overestimate their compute requirements for AI workloads. Before investing in GPU clusters, estimate your actual FLOPS requirements using published benchmarks and model scaling laws. Start with cloud-based GPU instances to validate your needs before committing to capital-intensive hardware purchases. Consider the total cost of ownership: hardware cost represents only 30–40% of the total cost of AI computing — power, cooling, networking, and personnel often exceed the hardware investment over a three-year period.

Explore mixed-precision and quantised computing to reduce FLOPS requirements by 50–75% with minimal accuracy loss. Most AI workloads do not require full 32-bit floating-point precision; 16-bit (FP16) or even 8-bit (INT8) arithmetic delivers equivalent model quality at dramatically lower computational cost. Also evaluate whether your AI use cases can be served by pre-trained models via API (paying per token rather than per FLOPS), which can be 5–10 times cheaper for low-to-medium volume workloads than running your own GPU infrastructure. Finally, monitor the emerging generation of AI-specific chips from startups like Cerebras, Groq, and SambaNova, which promise order-of-magnitude improvements in FLOPS-per-watt efficiency.